Thank you & Suggestion: Exploring Kaggle or platform-agnostic support

[hosmallming]

Thank you so much for developing the Colab version of whisperjav. It has been a great help for users who cannot run it locally to enjoy the convenience of generating subtitles.
I’ve noticed that the free version of Colab can be somewhat unstable. From the perspective of free cloud GPU resources, Kaggle seems to be a more ideal alternative. I was wondering if it might be possible to develop a 'universal' script that is compatible across various cloud platforms? Thank you again for your hard work!

[meizhong986]

Good suggestion!

I've just pushed a new version that supports both Kaggle and Google;
It runs parallel on Kaggle and single GPU on Google.

It is ready for alpha testing. I don't have bandwidth right now to test it. Would you give it a try and let me know.

What it does in this new one:

1. Platform detection: The detect_platform() logic is sued to to run on Colab (Google Drive) and Kaggle (Input/Output tabs).

2. GPU availability awareness:

   • Kaggle (2x T4): Automatically activates PARALLEL_MODE to transcribe Pass 1 and Pass 2 simultaneously.
   • Colab (1x GPU): Automatically falls back to sequential processing.

3. Access to the output files:

   • Auto ZIP bundling at the end of transcription, and translation steps.
   • Kaggle Users : Download ZIP notification.
   • Colab Users auto download trigger via google.colab.files.download.

[hosmallming]

No problem at all, I'd love to test it out! Just a heads-up that I’m in a different time zone, so I’ll get back to you with the results a bit later.

[meizhong986]

Here is the notebook --you can find it in the notebook folder:

https://github.com/meizhong986/WhisperJAV/blob/main/notebook/WhisperJAV_colab_parallel_expert.ipynb

You should see the icons for openning the notebook:

[hosmallming]

Tested 2 more videos with the same result: 'Merged: 0 + 0 → 0 subtitles'. (Logs and params attached below).

log_video_1.txt
log_video_2.txt

One notable observation: during execution, CPU and GPU1 showed high utilization, but GPU0 was idle.

This hardware behavior might be a helpful clue for your investigation.
Thanks for your hard work on this!

[meizhong986]

This is very helpful. Thank you!

It is now fixed in the new notebook release (commit). You'd need to open the notebook from the original link again to get the latest version (commit f083e5f). https://github.com/meizhong986/WhisperJAV/blob/main/notebook/WhisperJAV_colab_parallel_expert.ipynb

Please let me know how it goes. Thanks for your help.

What was the issue:

• Symptom: "CPU 100%, GPU 0 0%".
• Cause: The faster-whisper library (used in the balanced pipeline) could not find the NVIDIA cuBLAS/cuDNN libraries in the Kaggle environment. It silently fell back to running on the CPU, which explains the maxed-out CPU and idle GPU.

Why that happened:

The evolution of the notebook from a Colab-native script to a "Universal" one carried over some assumptions about the environment that don't hold true on Kaggle.

1. Correct PATH to the Libraries (LD_LIBRARY_PATH): On Colab, NVIDIA libraries (cuBLAS, cuDNN) are in standard paths. On Kaggle, they are installed via PyPI in site-packages/nvidia/*/lib and are not in the system search path by default. This caused faster-whisper (Pass 1) to silently fail GPU initialization and fall back to CPU.

2. Storage Limits (tmp vs /kaggle/working): Colab's tmp is flexible. Kaggle's tmp is a small RAM disk or limited partition. Writing heavy WAV/PCM files there caused them to be truncated (0 bytes), which explains why Pass 2 (transformers) ran on GPU but found "0 subtitles".

[hosmallming]

Hi, here’s an error log from the new version. Interestingly, while Pass 1 ran into an error, Pass 2 actually managed to generate the subtitles without any issues. Hope this information is helpful for your debugging. Thanks!

[meizhong986]

Thanks for the test. It seems that the previous fix didn't resolve Kaggle specific installation conflict with preinstalled components. This time the issue was matplotlib. Assuming that the conflict is an architectural dilemma, I refactored the installation so in Kaggle, whisperjav will rely solely on preinstalled packages. This will avoid the conflicts but needs testing, because there might be unexpected version mismatches. If this fails too, then I need to make the whisperjav to be standalone, installing all its necessary packages to override system preinstalls.

I have just pushed a new notebook commit. The new one has these fixes:

Kaggle: Refactored the entire INSTALLATION and setup blocks to use the new architecture.

1. Environment Locking: The notebook now runs pip freeze > constraints.txt first.

2. Constraint-Based Install: All new packages are installed with -c constraints.txt, forcing pip to respect the existing versions of system libraries (like numpy, matplotlib, torch) instead of upgrading them and breaking things.

3. Minimalist Dependency List: I removed all standard libraries (numpy, scipy, requests, matplotlib, tqdm, etc.) from the installation list. We now rely on the robust pre-installed versions provided by Kaggle/Colab.

4. Runtime Linker Configuration: The LD_LIBRARY_PATH logic (required for faster-whisper to see the correct CUDA libs) has been moved from a hack in the platform detection to a formal configure_cuda_runtime_paths() function that runs cleanly after installation.

Thanks for your help.

[meizhong986]

In the new test runs, would you be able to send me all output texts of all cells in the notebook --I wonder if Kaggle allows copy all or ctlr+A.

[hosmallming]

Attached are my test results. I’m not really a technical person, so please excuse my lack of expertise, but I hope these findings provide some value to the project. Thanks again!
log.txt

[meizhong986]

I just made new changes in the notebook. You can pick up the latest release from the original link.
This should have the final fix.
Would you run your test and send me the logs again.
Much appreciated!

[hosmallming]

Here is the log from the newest update. I appreciate your continued help with this project. Thank you!
log.txt

[meizhong986]

So the parallel GPUs now work. But something is odd with the SRT production. I need to work on it.

[hosmallming]

I was also surprised to see the subtitles missing from the output. Thanks again for your time and help in looking into this!

[meizhong986]

I just updated the notebook. I made gemini-3 fight it with gpt5.2 :)

This should:

• Fix the dual GPU both parallel for each pass/run;
• New pipeline presets: by default:
  pass 1 is balanced | aggressive | semantic | TEN | large-v2
  pass 2 is fidelity | balanced | semantic | TEN | turbo
• Every 8 seconds console updates the progress --prints out last 4 lines of the log
• All output artefacts are zipped for download including SRTs, and logs;
• Translator is still getting error --will fix in next step if the rest works;

Please give it a try and let me know.

[hosmallming]

It’s great to hear that Gemini 3 is now supported. Please find the latest execution log attached for your reference. Thanks for your continued help with this!
log.txt

[meizhong986]

There was a crash. I have now fixed it in the new notebook release.
Crash Fix: SciPy vs NumPy 2.0.

Please try the new version and let me know.

Change notes:

• Scipy > 1.14.0
• Internet Check: Switched to a multi-endpoint check (Baidu, Google, 1.1.1.1) to fix the "Internet unreachable" error for users in China.
• Timeouts: Added timeout=600 (10 minutes) to all subprocess.run calls to prevent hanging forever on slow downloads/installs.
• Dependency Safety: Added a final explicit install of numpy>=2.0 after the git packages to prevent version downgrades.
• Verification: Added must(shutil.which("ffmpeg")) to confirm FFmpeg installation.
• TEN VAD: Clarified the warning message to "TEN VAD Warning (Import failed - Silero fallback)".
• Cleanup: Removed dead code (FAIL_FAST check, unused imports), but kept WHISPERJAV_SETUP_COMPLETE as it is critically used in Step 3 for platform detection.

[hosmallming]

Please find the current log attached. It occurred to me that my suggestion may have added a lot to your workload lately, and I sincerely apologize for any inconvenience. I truly appreciate your dedication to this project!
log.txt

[meizhong986]

I made changes in whisperjav for that: scipy>=1.10.1 → scipy>=1.14.0;
Would you try the test again please.

[hosmallming]

It's working!
log.txt
VidTrim_6463_pass1.log
VidTrim_6463_pass2.log

[hosmallming]

Ran into an error when testing the translation function—log attached. Also,
I couldn't seem to find the language setting for translations in this build. Is there a specific way to change the target language now? Thanks for your help!
log.txt

[meizhong986]

Great. Thanks !
And, yes that is the next in todo list.

[meizhong986]

I just updated the translation section of the notebook. But I haven't tested it yet.
I hope it works out of the box.
Please let me know.
PS. The default target languages in the form are English, Chinese, Spanish, and Indonesian. Please let me what other languages might be needed in the short list. In this version I haven't done a free form for a user to just write up the target language --it will require more coding and mapping.

[hosmallming]

Hey! I made some small tweaks with Gemini's help to get the subtitles outputting in Chinese—it works great! OpenRouter is running perfectly, but I’m hitting a 429 error with Gemini because it’s still using the older gemini-2-flash. Would you consider switching it to gemini-3-flash? I’d love to test that out for you.
Batch translation is also working well. On another note, I ran a test on a Korean show (I love it!) and spotted a possible bug. I’ve attached the details for your reference. Really appreciate all the hard work you're putting in!
bug.txt
Runningman_pass2.log
Runningman_pass1.log

[meizhong986]

Thanks for testing. Much appreciated.

About the translation with gemini-3-flash in the notebook. Until I add it to the core whisperjav-translate, for now you can simply enter model name (e.g., gemini-3-flash) into the model_override field in the form.

About the error: I was surprised by that. It should not have happened. I just pushed a new commit. Please try the notebook again.

[hosmallming]

I tried both Japanese and Korean input, but the result is still the same. Hope this helps with your investigation! Thanks for all your help
log.txt

[meizhong986]

HHmm. I find Kaggle very unreliable. It seems that it pins different environments each time a notebook is created. That makes it very unreliable. I would imagine that there must be a way to pin it down to force it not to change environment. But my Kaggle expertise is lacking.

[hosmallming]

Following Gemini's suggestions, I modified the code and fortunately, it executed successfully! I have attached the detailed explanation for your reference.
report.txt

I also managed to generate Chinese subtitles using gemini-2.5-flash. However, I was surprised to get an error saying gemini-3-flash wasn't found—I suspect the model source might not be Google AI Studio?

After a few test runs, I've realized that running translations on Kaggle or Colab might not be the best idea. It burns through valuable GPU quotas and triggers API rate limits way too easily. It's probably better to just stick to transcription on these platforms.

[TinyRick1489]

You are right. I usually use Kaggle (with GPU) for transcription and the use Colab (without GPU) for translation. I could have used kaggle (without GPU) for translation but upload files to Colab is much easier than to Kaggle.

[meizhong986]

Exactly. It is a waste of GPU credits.

I'm taking a couple of days break from this thread and will be back in couple of days.

[TinyRick1489]

Take your time. Right now I am using the latest notebook and its working without any issues.

[meizhong986]

I actually tried to make the notebook to switch-off the GPU after transcription to continue to translation, but that required a kill os and restart --which is not a good user experience.

…

On Wed, Jan 14, 2026 at 12:53 PM TinyRick1489 ***@***.***> wrote: You are right. I usually use Kaggle (with GPU) for transcription and the use Colab (without GPU) for translation. I could have used kaggle (without GPU) for translation but upload files to Colab is much easier than to Kaggle. — Reply to this email directly, view it on GitHub <#121 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BDME2BUBJVHVFWZTLMTR5DL4GY35VAVCNFSM6AAAAACREP7IHGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBZGU3DQMI> . You are receiving this because you commented.Message ID: ***@***.***>

[hosmallming]

I noticed that a NumPy version warning was triggered while running the latest version of the program. I'm providing this for your reference to ensure compatibility. Thank you!
log.txt

[meizhong986]

I have updated parallel notebook to use install_colab.sh for Colab portion (while keeping direct pip for Kaggle which has different GPU setup). Please let me know if still any issues.

[hosmallming]

It works perfect! Thanks a lot!

[TinyRick1489]

Hi @meizhong986
I tested the notebook and it worked.
I wanted to use semantic scene detection but as there is no argument for it in CLI single mode, I can't use it.
In previous version of notebook you used pass1 for both GPU0 and GPU1 and with that we were able to change scene detection.
Can we use that approach again?

[meizhong986]

@TinyRick1489 , I just updated the notebook.
There was a bug where the Scene Detection choice was being ignored.
Now when you select 'Semantic' in the Expert settings, it will correctly force the use of the semantic scene detector as intended. Please let me know.

[TinyRick1489]

Temps are getting deleted when one of the pass is completed.
Using --keep-temp flag solves it.

[hosmallming]

Have you seen the new Qwen3-ASR series? Word is that they’re beating Whisper-large-v3 across multiple benchmarks. Thought it might be interesting to look into if you have the time! Really appreciate your constant effort on this project.

[hosmallming]

My current configurations are as follows:

Pass 1: transformers/silero/kotoba-v2.2

Pass 2: fidelity/ten/large-v2

Merge: prefer second pass

I’ve noticed a sync issue where the subtitles appear slightly ahead of the audio. Do you have any recommended settings or best practices to resolve this timing discrepancy? Thank you!

[meizhong986]

TEN has that characteristics --hence I think the source is in your Pass 2.
I need to adjust the TEN settings and post adjustments.
Open an issue so I won't forget :)

[hosmallming]

I wanted to share two practical tips from my recent tests.
First, I fixed an issue with certain videos that wouldn't load by using -f mpegts in FFmpeg, as they were actually TS streams with an MP4 extension.
Second, I’ve found that mounting Google Drive in Kaggle is a lifesaver. Even if it feels like an extra step, it bypasses Kaggle's strict filename limits and makes organizing files much easier.
Hope this helps with the next version!
Thanks for your hard work.

[meizhong986]

Thanks for the good suggestions. I didn't know Kaggle had google drive connector. Can it also save / write to drive or only reads?

[hosmallming]

I'm using PyDrive2 to handle both reading and writing to Google Drive.

[meizhong986]

Great suggestion. If you don't mind, please add this into the issues as feature suggestion. That way I grab it during the next iteration and work on it --so it doesn't get forgotten. Thanks

[hosmallming]

I’m so happy to see whisperjav making such significant progress! Whenever you have a spare moment, would you mind updating the Kaggle version as well? Thank you so much for your amazing work!